|
researchaudio.io
Issue 75
·
June 22, 2026
|
The person prompting Claude Code is the bottleneck. The loop is no longer optional.
The 2026 workflow is not a prompt. It is five primitives wired into a system you barely watch.
Boris Cherny the person who built Claude Code said something at a private WorkOS event on June 2 that broke containment within a week:
|
"I don't prompt Claude anymore. I have loops that are running. They're the ones that are prompting Claude and figuring out what to do. My job is to write loops."
Boris Cherny, WorkOS event, June 2 2026
|
Peter Steinberger said almost the same thing the same week: You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents.
Both said this in June 2026. Both said it without coordinating. Both said it because the workflow underneath them had quietly crossed a line you can now hand the prompt to a small system that runs the agent on a schedule, inspects the result, updates a memory file, and starts the next pass without you.
That small system is the loop. It is not a clever hack. It is the new abstraction layer between you and the agent, and most engineering teams are not building it yet. They are still hand-prompting Claude one turn at a time and wondering why their 4x context-window competitors feel like they have three engineers working through the night.
|
June 2026, two quotes, same week
"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."
Peter Steinberger
"I don't prompt Claude anymore. I have loops running that prompt Claude. My job is to write loops."
Boris Cherny, head of Claude Code, Anthropic
|
What "loop engineering" actually means
In plain English: A loop is a tiny program that runs an agent on its own, decides what to do next based on what the agent did, and remembers the answer for next time. You write the loop once. The agent runs it. You read the output.
The term was coined into common use by Addy Osmani's June 7 post Loop Engineering: Designing loops that prompt coding agents (addyosmani.com/blog/loop-engineering/). His five-piece definition is the cleanest version in circulation:
|
Primitive
|
What it does
|
| Automations | Run the agent on a schedule |
| Worktrees | Isolate parallel agents |
| Skills | Codify project knowledge |
| Connectors | MCP into Slack, Linear, your DB |
| Sub-agents | Maker writes, checker grades |
| State | Memory outside the conversation |
Notice the table. The names are different. The capability is the same. Addy writes:
|
"Once you notice the shape is the same you stop arguing about which tool, you just design a loop that still works no matter which one you happen to be sitting in."
|
That is the punchline nobody wanted to hear. The agent stopped being the bottleneck. The loop is.
AI World Brief week of June 15, 2026
| Boris Cherny confirmed at WorkOS: agents now run on user-written loops, not user-written prompts. |
| Databricks open-sourced Omnigent (Apache 2.0), a meta-harness wrapping Claude Code, Codex, Pi under one uniform API. |
| Addy Osmani published the canonical loop engineering playbook. 10 points on HN, 6 comments in 24 hours. |
| Show HN: Millrace (51 stars) first framework to ship "governed loops" where the prompt mutates based on prior step success. |
| Show HN: Loopy (9 stars) local launchd watcher that mines your Claude Code transcripts and proposes new loops. |
The five primitives, mapped
Anthropic Claude Code · OpenAI Codex app
|
Figure 1: Five primitives, two products, identical shape. Source: addyosmani.com/blog/loop-engineering/
|
How a single loop looks, end to end
In plain English: You write one prompt that calls a skill. That skill reads yesterday's failures. For each real problem, a worker agent opens its own copy of the repo, writes a fix, and a separate reviewer agent grades it. The loop files the PR and updates the ticket. Tomorrow morning the same loop runs again and picks up where yesterday stopped.
|
06:00 Automation wakes up. Runs "triage-yesterday" skill.
Reads CI failures, open issues, last 50 commits.
Writes findings to state.md.
06:04 For each finding marked "worth doing":
- open a git worktree (branch: fix/)
- spawn sub-agent A: implementer (Sonnet, high effort)
- spawn sub-agent B: reviewer (Opus, low effort, diff only)
- if B rejects: state.md gets the reason, retry budget = 2
06:31 Loop runs /goal with stop condition:
"all tests in test/auth pass AND lint is clean"
07:15 Loop opens PR via the GitHub MCP connector.
Updates the Linear ticket.
Posts to #eng-bots: "PR #4821 ready for review"
07:16 state.md: done=3, blocked=1, ignored=12
07:17 Loop exits. You wake up. You read state.md. You review 3 PRs.
|
You wrote none of those steps. The agents wrote the code. The reviewer graded it. The connectors filed the work. You are the bottleneck, not the prompt, not the model, not the context window. You are.
That is the trade. Boris is not saying his job got easier. He is saying the leverage point moved.
The numbers nobody is publishing
% of loop runs that actually do useful work
|
Figure 2: Field estimates from Boris (20-30 PRs/day), Omnigent (45-55% cross-vendor catch), Loopy docs. Not lab numbers.
|
Community Reaction
Loop engineering is a personal blog post and a private event quote. There is no company blog. There is no product launch. That is why the strongest reactions are in the comments, not the press.
|
"I just don't see any way you can work like this and maintain comprehension of the system being built? Perhaps for toy projects or research, but for a production system you're accountable for understanding, maintaining and continuing to develop?"
aocallaghan17 · HN, June 13 2026
|
|
"If the future is 'loop engineering' then you will be burning through tokens and serious cash using cloud models."
jborak · HN, June 15 2026, 4x RTX 5070s locally
|
|
"Loopy runs quietly in the background, watches session transcripts, recognizes patterns and proposes ready-to-install automation loops. Right now, the trend is on loop engineering and I think this is one that'll stay for sometime."
secretbuilds · HN, June 14 2026, github.com/secretbuilds/loopy
|
|
"Omnigent sits above Pi, Claude, Codex, etc. and wraps each in a uniform API. Approval and cost policies enforced at the harness layer rather than living in a prompt the agent can talk its way around."
saj1th · HN, June 13 2026
|
|
"Two people can build the exact same loop and get completely opposite results. One uses it to move faster on work they understand deeply. The other uses it to avoid understanding the work at all. The loop doesn't know the difference. You do."
Addy Osmani · Loop Engineering, June 7 2026
|
Where the leverage moved
The bottleneck, by year
|
Figure 3: Same conversation, different constraint, every 9-12 months. Source: Boris Cherny (WorkOS, June 2 2026), Addy Osmani (June 7 2026), community consensus HN June 9-16.
|
What This Means for Engineers
|
Juniors
Don't wait for permission to write a loop. Find one recurring task you keep meaning to do dependency bumps, lint cleanup, changelog drift and turn it into a /loop cron.
|
|
Seniors
Your value is no longer the prompt you wrote. It is the loop you shipped that 5 other engineers can read. Stop hoarding clever prompts. Write the loop, write the skill, write the AGENTS.md.
|
|
Hiring managers
Stop hiring prompt engineers. Hire loop engineers. Same as the sysadmin-to-SRE jump: one writes the script, the other runs the system that runs the scripts.
|
|
Founders
The window is open right now for two products and closing fast. (1) Loop observability token-cost dashboards, sub-agent success rates. (2) Cross-vendor maker/checker infrastructure. Omnigent took shot two. Shot one is wide open.
|
|
Share this
The 2026 workflow is not a prompt. It is five primitives wired into a system you barely watch.
Boris Cherny · Peter Steinberger · Addy Osmani
|
|
"The loop doesn't know the difference. You do. That's what makes loop design harder than prompt engineering, not easier. Cherny's point isn't that the work got easier. It's that the leverage point moved."
Addy Osmani, Loop Engineering, June 7 2026
|
Reader Challenge
| 1. Pick one task you do by hand every week. Write a SKILL.md for it. Wire it to a /loop cron. Measure how many times you actually look at the output in 30 days. |
| 2. Take your last 5 PRs. For each, write a sub-agent prompt that would have caught the review comment you left. That is your first verifier skill. |
| 3. Honest version: how many lines of your repo could you explain in 60 seconds without grep? If the number is dropping, you are accumulating comprehension load. |
Next Issue
Issue 76: The maker/checker split is the actual breakthrough. Three reasons every solo developer using one model is going to fall behind a team using two.
|
